fix(remote): never auto-revive a remote session after a clean agent exit

[timkjr]

Problem

A normal ctrl-c / ctrl-d / exit inside a remote CLI session (Claude, OpenCode, OMP, …) auto-respawns a fresh agent. The COD-108 remote-reconnect watcher polls every 5s and, for any remote session whose local pane is dead, emits remoteSessionDropped — which reattaches by re-running the pane command.

The watcher intended to recover from a transport drop (ssh drops, agent still running on the remote). But isPaneDead() is true in BOTH cases: a transport drop AND a normal agent exit. A clean exit tears down the durable remote tmux session (codeman-ssh-*, remain-on-exit failed destroys the session when its only pane exits), so the watcher cannot tell the two apart — and re-runs the command, launching a brand-new conversation.

Claude only looked okay: its remote launch is claude --session-id <id> || claude --resume <id>, so the fresh run resumed — but not before printing a loud "already in use" error. OpenCode / OMP started fresh every time.

Fix

Only auto-reconnect when the durable remote tmux session is verifiably still alive on the remote host:

• remoteTmuxSessionAlive() runs tmux -L codeman-remote has-session -t codeman-ssh-<id> over ssh (exit 0 = alive).
• decideReconnect() gains a remote-gone skip: when the remote session is gone (clean exit → do not revive) OR the probe is unknown (unreachable host → fail closed, do not revive), the watcher does nothing.
• The probe is cached per-session and fired async, so the sync 5s tick never blocks on an ssh round-trip; a clean exit flips the cache to false and the auto-revive stops.

Transport-drop behavior is unchanged: remote tmux alive → reconnect as before.

Testing

• 3 new unit cases in remote-auto-reconnect.test.ts pinning the decision: remote alive → emit; remote gone → skip; remote unknown → skip.
• Verified live: after ctrl-c / ctrl-d on remote Claude, OpenCode and OMP sessions, the pane stays dead (no auto-respawn).

This is intentionally independent of any CLI-specific resume logic — it fixes the watcher for every remote mode at once.

---

Note: I have a separate open PR (#353) for OMP backend support. This fix is deliberately scoped away from that branch; it applies to upstream master standalone.

[Ark0N]

Merged, thank you. The clean-exit rule is exactly right and it is what ships. One correction went in right behind it, so you know what changed: tmux has-session prints nothing on success (measured on a scratch socket: exit 0, empty stdout, the "can't find session" line goes to stderr), so stdout.trim().length > 0 read every live remote session as gone, and transport drops stopped reconnecting along with the clean exits. The follow-up (28b4423) classifies the probe by exit status instead (0 alive, ssh's 255 or a timeout unknown, anything else gone), forgets the cached answer whenever the pane is seen alive again so a stale true from one transport drop cannot revive the next clean exit, caps the probe at one in flight per session, and pins the command and the reattach-then-clean-exit sequence in tests. Ships in today's release.